Erasing dielectric breakdown artifacts to machine-learn charged Pt–water interfaces

Abstract

The recently introduced “response analysis in z-orientation” (RAZOR) model utilizes perturbation theory to machine learn the energy and force response to applied bias charges in atomistic simulations of electrified interfaces. While we have shown that RAZOR successfully reproduces ab initio results for adsorbates on metallic surfaces in an implicit solvent environment, real electrochemical applications require the inclusion of at least a few explicit H2O layers of the electrolyte. Here, we benchmark RAZOR’s performance in the description of the Pt(111)–H2O interface. We show that RAZOR can reliably reproduce ab initio molecular dynamics findings for bias-induced changes in H2O density and orientation profiles, as well as changes in the potential. To do so, a specific training procedure needs to be applied, which avoids the erroneous learning of artifacts from the dielectric breakdown of interfacial water and the accompanying charge transfer into the H2O layers. Ultimately, we are confident that RAZOR can provide quantitative predictions in a ±20 μC cm−2 window around the neutral-charged cell, which is sufficient for many pressing electrochemical questions.